sys/contrib/openzfs/module/zfs/vdev_initialize.c

   1 /*
   2  * CDDL HEADER START
   3  *
   4  * The contents of this file are subject to the terms of the
   5  * Common Development and Distribution License (the "License").
   6  * You may not use this file except in compliance with the License.
   7  *
   8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
   9  * or http://www.opensolaris.org/os/licensing.
  10  * See the License for the specific language governing permissions
  11  * and limitations under the License.
  12  *
  13  * When distributing Covered Code, include this CDDL HEADER in each
  14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
  15  * If applicable, add the following below this CDDL HEADER, with the
  16  * fields enclosed by brackets "[]" replaced with your own identifying
  17  * information: Portions Copyright [yyyy] [name of copyright owner]
  18  *
  19  * CDDL HEADER END
  20  */
  21
  22 /*
  23  * Copyright (c) 2016, 2019 by Delphix. All rights reserved.
  24  */
  25
  26 #include <sys/spa.h>
  27 #include <sys/spa_impl.h>
  28 #include <sys/txg.h>
  29 #include <sys/vdev_impl.h>
  30 #include <sys/metaslab_impl.h>
  31 #include <sys/dsl_synctask.h>
  32 #include <sys/zap.h>
  33 #include <sys/dmu_tx.h>
  34 #include <sys/vdev_initialize.h>
  35
  36 /*
  37  * Value that is written to disk during initialization.
  38  */
  39 #ifdef _ILP32
  40 unsigned long zfs_initialize_value = 0xdeadbeefUL;
  41 #else
  42 unsigned long zfs_initialize_value = 0xdeadbeefdeadbeeeULL;
  43 #endif
  44
  45 /* maximum number of I/Os outstanding per leaf vdev */
  46 int zfs_initialize_limit = 1;
  47
  48 /* size of initializing writes; default 1MiB, see zfs_remove_max_segment */
  49 unsigned long zfs_initialize_chunk_size = 1024 * 1024;
  50
  51 static boolean_t
  52 vdev_initialize_should_stop(vdev_t *vd)
  53 {
  54         return (vd->vdev_initialize_exit_wanted || !vdev_writeable(vd) ||
  55             vd->vdev_detached || vd->vdev_top->vdev_removing);
  56 }
  57
  58 static void
  59 vdev_initialize_zap_update_sync(void *arg, dmu_tx_t *tx)
  60 {
  61         /*
  62          * We pass in the guid instead of the vdev_t since the vdev may
  63          * have been freed prior to the sync task being processed. This
  64          * happens when a vdev is detached as we call spa_config_vdev_exit(),
  65          * stop the initializing thread, schedule the sync task, and free
  66          * the vdev. Later when the scheduled sync task is invoked, it would
  67          * find that the vdev has been freed.
  68          */
  69         uint64_t guid = *(uint64_t *)arg;
  70         uint64_t txg = dmu_tx_get_txg(tx);
  71         kmem_free(arg, sizeof (uint64_t));
  72
  73         vdev_t *vd = spa_lookup_by_guid(tx->tx_pool->dp_spa, guid, B_FALSE);
  74         if (vd == NULL || vd->vdev_top->vdev_removing || !vdev_is_concrete(vd))
  75                 return;
  76
  77         uint64_t last_offset = vd->vdev_initialize_offset[txg & TXG_MASK];
  78         vd->vdev_initialize_offset[txg & TXG_MASK] = 0;
  79
  80         VERIFY(vd->vdev_leaf_zap != 0);
  81
  82         objset_t *mos = vd->vdev_spa->spa_meta_objset;
  83
  84         if (last_offset > 0) {
  85                 vd->vdev_initialize_last_offset = last_offset;
  86                 VERIFY0(zap_update(mos, vd->vdev_leaf_zap,
  87                     VDEV_LEAF_ZAP_INITIALIZE_LAST_OFFSET,
  88                     sizeof (last_offset), 1, &last_offset, tx));
  89         }
  90         if (vd->vdev_initialize_action_time > 0) {
  91                 uint64_t val = (uint64_t)vd->vdev_initialize_action_time;
  92                 VERIFY0(zap_update(mos, vd->vdev_leaf_zap,
  93                     VDEV_LEAF_ZAP_INITIALIZE_ACTION_TIME, sizeof (val),
  94                     1, &val, tx));
  95         }
  96
  97         uint64_t initialize_state = vd->vdev_initialize_state;
  98         VERIFY0(zap_update(mos, vd->vdev_leaf_zap,
  99             VDEV_LEAF_ZAP_INITIALIZE_STATE, sizeof (initialize_state), 1,
 100             &initialize_state, tx));
 101 }
 102
 103 static void
 104 vdev_initialize_change_state(vdev_t *vd, vdev_initializing_state_t new_state)
 105 {
 106         ASSERT(MUTEX_HELD(&vd->vdev_initialize_lock));
 107         spa_t *spa = vd->vdev_spa;
 108
 109         if (new_state == vd->vdev_initialize_state)
 110                 return;
 111
 112         /*
 113          * Copy the vd's guid, this will be freed by the sync task.
 114          */
 115         uint64_t *guid = kmem_zalloc(sizeof (uint64_t), KM_SLEEP);
 116         *guid = vd->vdev_guid;
 117
 118         /*
 119          * If we're suspending, then preserving the original start time.
 120          */
 121         if (vd->vdev_initialize_state != VDEV_INITIALIZE_SUSPENDED) {
 122                 vd->vdev_initialize_action_time = gethrestime_sec();
 123         }
 124         vd->vdev_initialize_state = new_state;
 125
 126         dmu_tx_t *tx = dmu_tx_create_dd(spa_get_dsl(spa)->dp_mos_dir);
 127         VERIFY0(dmu_tx_assign(tx, TXG_WAIT));
 128         dsl_sync_task_nowait(spa_get_dsl(spa), vdev_initialize_zap_update_sync,
 129             guid, tx);
 130
 131         switch (new_state) {
 132         case VDEV_INITIALIZE_ACTIVE:
 133                 spa_history_log_internal(spa, "initialize", tx,
 134                     "vdev=%s activated", vd->vdev_path);
 135                 break;
 136         case VDEV_INITIALIZE_SUSPENDED:
 137                 spa_history_log_internal(spa, "initialize", tx,
 138                     "vdev=%s suspended", vd->vdev_path);
 139                 break;
 140         case VDEV_INITIALIZE_CANCELED:
 141                 spa_history_log_internal(spa, "initialize", tx,
 142                     "vdev=%s canceled", vd->vdev_path);
 143                 break;
 144         case VDEV_INITIALIZE_COMPLETE:
 145                 spa_history_log_internal(spa, "initialize", tx,
 146                     "vdev=%s complete", vd->vdev_path);
 147                 break;
 148         default:
 149                 panic("invalid state %llu", (unsigned long long)new_state);
 150         }
 151
 152         dmu_tx_commit(tx);
 153
 154         if (new_state != VDEV_INITIALIZE_ACTIVE)
 155                 spa_notify_waiters(spa);
 156 }
 157
 158 static void
 159 vdev_initialize_cb(zio_t *zio)
 160 {
 161         vdev_t *vd = zio->io_vd;
 162         mutex_enter(&vd->vdev_initialize_io_lock);
 163         if (zio->io_error == ENXIO && !vdev_writeable(vd)) {
 164                 /*
 165                  * The I/O failed because the vdev was unavailable; roll the
 166                  * last offset back. (This works because spa_sync waits on
 167                  * spa_txg_zio before it runs sync tasks.)
 168                  */
 169                 uint64_t *off =
 170                     &vd->vdev_initialize_offset[zio->io_txg & TXG_MASK];
 171                 *off = MIN(*off, zio->io_offset);
 172         } else {
 173                 /*
 174                  * Since initializing is best-effort, we ignore I/O errors and
 175                  * rely on vdev_probe to determine if the errors are more
 176                  * critical.
 177                  */
 178                 if (zio->io_error != 0)
 179                         vd->vdev_stat.vs_initialize_errors++;
 180
 181                 vd->vdev_initialize_bytes_done += zio->io_orig_size;
 182         }
 183         ASSERT3U(vd->vdev_initialize_inflight, >, 0);
 184         vd->vdev_initialize_inflight--;
 185         cv_broadcast(&vd->vdev_initialize_io_cv);
 186         mutex_exit(&vd->vdev_initialize_io_lock);
 187
 188         spa_config_exit(vd->vdev_spa, SCL_STATE_ALL, vd);
 189 }
 190
 191 /* Takes care of physical writing and limiting # of concurrent ZIOs. */
 192 static int
 193 vdev_initialize_write(vdev_t *vd, uint64_t start, uint64_t size, abd_t *data)
 194 {
 195         spa_t *spa = vd->vdev_spa;
 196
 197         /* Limit inflight initializing I/Os */
 198         mutex_enter(&vd->vdev_initialize_io_lock);
 199         while (vd->vdev_initialize_inflight >= zfs_initialize_limit) {
 200                 cv_wait(&vd->vdev_initialize_io_cv,
 201                     &vd->vdev_initialize_io_lock);
 202         }
 203         vd->vdev_initialize_inflight++;
 204         mutex_exit(&vd->vdev_initialize_io_lock);
 205
 206         dmu_tx_t *tx = dmu_tx_create_dd(spa_get_dsl(spa)->dp_mos_dir);
 207         VERIFY0(dmu_tx_assign(tx, TXG_WAIT));
 208         uint64_t txg = dmu_tx_get_txg(tx);
 209
 210         spa_config_enter(spa, SCL_STATE_ALL, vd, RW_READER);
 211         mutex_enter(&vd->vdev_initialize_lock);
 212
 213         if (vd->vdev_initialize_offset[txg & TXG_MASK] == 0) {
 214                 uint64_t *guid = kmem_zalloc(sizeof (uint64_t), KM_SLEEP);
 215                 *guid = vd->vdev_guid;
 216
 217                 /* This is the first write of this txg. */
 218                 dsl_sync_task_nowait(spa_get_dsl(spa),
 219                     vdev_initialize_zap_update_sync, guid, tx);
 220         }
 221
 222         /*
 223          * We know the vdev struct will still be around since all
 224          * consumers of vdev_free must stop the initialization first.
 225          */
 226         if (vdev_initialize_should_stop(vd)) {
 227                 mutex_enter(&vd->vdev_initialize_io_lock);
 228                 ASSERT3U(vd->vdev_initialize_inflight, >, 0);
 229                 vd->vdev_initialize_inflight--;
 230                 mutex_exit(&vd->vdev_initialize_io_lock);
 231                 spa_config_exit(vd->vdev_spa, SCL_STATE_ALL, vd);
 232                 mutex_exit(&vd->vdev_initialize_lock);
 233                 dmu_tx_commit(tx);
 234                 return (SET_ERROR(EINTR));
 235         }
 236         mutex_exit(&vd->vdev_initialize_lock);
 237
 238         vd->vdev_initialize_offset[txg & TXG_MASK] = start + size;
 239         zio_nowait(zio_write_phys(spa->spa_txg_zio[txg & TXG_MASK], vd, start,
 240             size, data, ZIO_CHECKSUM_OFF, vdev_initialize_cb, NULL,
 241             ZIO_PRIORITY_INITIALIZING, ZIO_FLAG_CANFAIL, B_FALSE));
 242         /* vdev_initialize_cb releases SCL_STATE_ALL */
 243
 244         dmu_tx_commit(tx);
 245
 246         return (0);
 247 }
 248
 249 /*
 250  * Callback to fill each ABD chunk with zfs_initialize_value. len must be
 251  * divisible by sizeof (uint64_t), and buf must be 8-byte aligned. The ABD
 252  * allocation will guarantee these for us.
 253  */
 254 /* ARGSUSED */
 255 static int
 256 vdev_initialize_block_fill(void *buf, size_t len, void *unused)
 257 {
 258         ASSERT0(len % sizeof (uint64_t));
 259 #ifdef _ILP32
 260         for (uint64_t i = 0; i < len; i += sizeof (uint32_t)) {
 261                 *(uint32_t *)((char *)(buf) + i) = zfs_initialize_value;
 262         }
 263 #else
 264         for (uint64_t i = 0; i < len; i += sizeof (uint64_t)) {
 265                 *(uint64_t *)((char *)(buf) + i) = zfs_initialize_value;
 266         }
 267 #endif
 268         return (0);
 269 }
 270
 271 static abd_t *
 272 vdev_initialize_block_alloc(void)
 273 {
 274         /* Allocate ABD for filler data */
 275         abd_t *data = abd_alloc_for_io(zfs_initialize_chunk_size, B_FALSE);
 276
 277         ASSERT0(zfs_initialize_chunk_size % sizeof (uint64_t));
 278         (void) abd_iterate_func(data, 0, zfs_initialize_chunk_size,
 279             vdev_initialize_block_fill, NULL);
 280
 281         return (data);
 282 }
 283
 284 static void
 285 vdev_initialize_block_free(abd_t *data)
 286 {
 287         abd_free(data);
 288 }
 289
 290 static int
 291 vdev_initialize_ranges(vdev_t *vd, abd_t *data)
 292 {
 293         range_tree_t *rt = vd->vdev_initialize_tree;
 294         zfs_btree_t *bt = &rt->rt_root;
 295         zfs_btree_index_t where;
 296
 297         for (range_seg_t *rs = zfs_btree_first(bt, &where); rs != NULL;
 298             rs = zfs_btree_next(bt, &where, &where)) {
 299                 uint64_t size = rs_get_end(rs, rt) - rs_get_start(rs, rt);
 300
 301                 /* Split range into legally-sized physical chunks */
 302                 uint64_t writes_required =
 303                     ((size - 1) / zfs_initialize_chunk_size) + 1;
 304
 305                 for (uint64_t w = 0; w < writes_required; w++) {
 306                         int error;
 307
 308                         error = vdev_initialize_write(vd,
 309                             VDEV_LABEL_START_SIZE + rs_get_start(rs, rt) +
 310                             (w * zfs_initialize_chunk_size),
 311                             MIN(size - (w * zfs_initialize_chunk_size),
 312                             zfs_initialize_chunk_size), data);
 313                         if (error != 0)
 314                                 return (error);
 315                 }
 316         }
 317         return (0);
 318 }
 319
 320 static void
 321 vdev_initialize_calculate_progress(vdev_t *vd)
 322 {
 323         ASSERT(spa_config_held(vd->vdev_spa, SCL_CONFIG, RW_READER) ||
 324             spa_config_held(vd->vdev_spa, SCL_CONFIG, RW_WRITER));
 325         ASSERT(vd->vdev_leaf_zap != 0);
 326
 327         vd->vdev_initialize_bytes_est = 0;
 328         vd->vdev_initialize_bytes_done = 0;
 329
 330         for (uint64_t i = 0; i < vd->vdev_top->vdev_ms_count; i++) {
 331                 metaslab_t *msp = vd->vdev_top->vdev_ms[i];
 332                 mutex_enter(&msp->ms_lock);
 333
 334                 uint64_t ms_free = msp->ms_size -
 335                     metaslab_allocated_space(msp);
 336
 337                 if (vd->vdev_top->vdev_ops == &vdev_raidz_ops)
 338                         ms_free /= vd->vdev_top->vdev_children;
 339
 340                 /*
 341                  * Convert the metaslab range to a physical range
 342                  * on our vdev. We use this to determine if we are
 343                  * in the middle of this metaslab range.
 344                  */
 345                 range_seg64_t logical_rs, physical_rs;
 346                 logical_rs.rs_start = msp->ms_start;
 347                 logical_rs.rs_end = msp->ms_start + msp->ms_size;
 348                 vdev_xlate(vd, &logical_rs, &physical_rs);
 349
 350                 if (vd->vdev_initialize_last_offset <= physical_rs.rs_start) {
 351                         vd->vdev_initialize_bytes_est += ms_free;
 352                         mutex_exit(&msp->ms_lock);
 353                         continue;
 354                 } else if (vd->vdev_initialize_last_offset >
 355                     physical_rs.rs_end) {
 356                         vd->vdev_initialize_bytes_done += ms_free;
 357                         vd->vdev_initialize_bytes_est += ms_free;
 358                         mutex_exit(&msp->ms_lock);
 359                         continue;
 360                 }
 361
 362                 /*
 363                  * If we get here, we're in the middle of initializing this
 364                  * metaslab. Load it and walk the free tree for more accurate
 365                  * progress estimation.
 366                  */
 367                 VERIFY0(metaslab_load(msp));
 368
 369                 zfs_btree_index_t where;
 370                 range_tree_t *rt = msp->ms_allocatable;
 371                 for (range_seg_t *rs =
 372                     zfs_btree_first(&rt->rt_root, &where); rs;
 373                     rs = zfs_btree_next(&rt->rt_root, &where,
 374                     &where)) {
 375                         logical_rs.rs_start = rs_get_start(rs, rt);
 376                         logical_rs.rs_end = rs_get_end(rs, rt);
 377                         vdev_xlate(vd, &logical_rs, &physical_rs);
 378
 379                         uint64_t size = physical_rs.rs_end -
 380                             physical_rs.rs_start;
 381                         vd->vdev_initialize_bytes_est += size;
 382                         if (vd->vdev_initialize_last_offset >
 383                             physical_rs.rs_end) {
 384                                 vd->vdev_initialize_bytes_done += size;
 385                         } else if (vd->vdev_initialize_last_offset >
 386                             physical_rs.rs_start &&
 387                             vd->vdev_initialize_last_offset <
 388                             physical_rs.rs_end) {
 389                                 vd->vdev_initialize_bytes_done +=
 390                                     vd->vdev_initialize_last_offset -
 391                                     physical_rs.rs_start;
 392                         }
 393                 }
 394                 mutex_exit(&msp->ms_lock);
 395         }
 396 }
 397
 398 static int
 399 vdev_initialize_load(vdev_t *vd)
 400 {
 401         int err = 0;
 402         ASSERT(spa_config_held(vd->vdev_spa, SCL_CONFIG, RW_READER) ||
 403             spa_config_held(vd->vdev_spa, SCL_CONFIG, RW_WRITER));
 404         ASSERT(vd->vdev_leaf_zap != 0);
 405
 406         if (vd->vdev_initialize_state == VDEV_INITIALIZE_ACTIVE ||
 407             vd->vdev_initialize_state == VDEV_INITIALIZE_SUSPENDED) {
 408                 err = zap_lookup(vd->vdev_spa->spa_meta_objset,
 409                     vd->vdev_leaf_zap, VDEV_LEAF_ZAP_INITIALIZE_LAST_OFFSET,
 410                     sizeof (vd->vdev_initialize_last_offset), 1,
 411                     &vd->vdev_initialize_last_offset);
 412                 if (err == ENOENT) {
 413                         vd->vdev_initialize_last_offset = 0;
 414                         err = 0;
 415                 }
 416         }
 417
 418         vdev_initialize_calculate_progress(vd);
 419         return (err);
 420 }
 421
 422 /*
 423  * Convert the logical range into a physical range and add it to our
 424  * avl tree.
 425  */
 426 static void
 427 vdev_initialize_range_add(void *arg, uint64_t start, uint64_t size)
 428 {
 429         vdev_t *vd = arg;
 430         range_seg64_t logical_rs, physical_rs;
 431         logical_rs.rs_start = start;
 432         logical_rs.rs_end = start + size;
 433
 434         ASSERT(vd->vdev_ops->vdev_op_leaf);
 435         vdev_xlate(vd, &logical_rs, &physical_rs);
 436
 437         IMPLY(vd->vdev_top == vd,
 438             logical_rs.rs_start == physical_rs.rs_start);
 439         IMPLY(vd->vdev_top == vd,
 440             logical_rs.rs_end == physical_rs.rs_end);
 441
 442         /* Only add segments that we have not visited yet */
 443         if (physical_rs.rs_end <= vd->vdev_initialize_last_offset)
 444                 return;
 445
 446         /* Pick up where we left off mid-range. */
 447         if (vd->vdev_initialize_last_offset > physical_rs.rs_start) {
 448                 zfs_dbgmsg("range write: vd %s changed (%llu, %llu) to "
 449                     "(%llu, %llu)", vd->vdev_path,
 450                     (u_longlong_t)physical_rs.rs_start,
 451                     (u_longlong_t)physical_rs.rs_end,
 452                     (u_longlong_t)vd->vdev_initialize_last_offset,
 453                     (u_longlong_t)physical_rs.rs_end);
 454                 ASSERT3U(physical_rs.rs_end, >,
 455                     vd->vdev_initialize_last_offset);
 456                 physical_rs.rs_start = vd->vdev_initialize_last_offset;
 457         }
 458         ASSERT3U(physical_rs.rs_end, >=, physical_rs.rs_start);
 459
 460         /*
 461          * With raidz, it's possible that the logical range does not live on
 462          * this leaf vdev. We only add the physical range to this vdev's if it
 463          * has a length greater than 0.
 464          */
 465         if (physical_rs.rs_end > physical_rs.rs_start) {
 466                 range_tree_add(vd->vdev_initialize_tree, physical_rs.rs_start,
 467                     physical_rs.rs_end - physical_rs.rs_start);
 468         } else {
 469                 ASSERT3U(physical_rs.rs_end, ==, physical_rs.rs_start);
 470         }
 471 }
 472
 473 static void
 474 vdev_initialize_thread(void *arg)
 475 {
 476         vdev_t *vd = arg;
 477         spa_t *spa = vd->vdev_spa;
 478         int error = 0;
 479         uint64_t ms_count = 0;
 480
 481         ASSERT(vdev_is_concrete(vd));
 482         spa_config_enter(spa, SCL_CONFIG, FTAG, RW_READER);
 483
 484         vd->vdev_initialize_last_offset = 0;
 485         VERIFY0(vdev_initialize_load(vd));
 486
 487         abd_t *deadbeef = vdev_initialize_block_alloc();
 488
 489         vd->vdev_initialize_tree = range_tree_create(NULL, RANGE_SEG64, NULL,
 490             0, 0);
 491
 492         for (uint64_t i = 0; !vd->vdev_detached &&
 493             i < vd->vdev_top->vdev_ms_count; i++) {
 494                 metaslab_t *msp = vd->vdev_top->vdev_ms[i];
 495                 boolean_t unload_when_done = B_FALSE;
 496
 497                 /*
 498                  * If we've expanded the top-level vdev or it's our
 499                  * first pass, calculate our progress.
 500                  */
 501                 if (vd->vdev_top->vdev_ms_count != ms_count) {
 502                         vdev_initialize_calculate_progress(vd);
 503                         ms_count = vd->vdev_top->vdev_ms_count;
 504                 }
 505
 506                 spa_config_exit(spa, SCL_CONFIG, FTAG);
 507                 metaslab_disable(msp);
 508                 mutex_enter(&msp->ms_lock);
 509                 if (!msp->ms_loaded && !msp->ms_loading)
 510                         unload_when_done = B_TRUE;
 511                 VERIFY0(metaslab_load(msp));
 512
 513                 range_tree_walk(msp->ms_allocatable, vdev_initialize_range_add,
 514                     vd);
 515                 mutex_exit(&msp->ms_lock);
 516
 517                 error = vdev_initialize_ranges(vd, deadbeef);
 518                 metaslab_enable(msp, B_TRUE, unload_when_done);
 519                 spa_config_enter(spa, SCL_CONFIG, FTAG, RW_READER);
 520
 521                 range_tree_vacate(vd->vdev_initialize_tree, NULL, NULL);
 522                 if (error != 0)
 523                         break;
 524         }
 525
 526         spa_config_exit(spa, SCL_CONFIG, FTAG);
 527         mutex_enter(&vd->vdev_initialize_io_lock);
 528         while (vd->vdev_initialize_inflight > 0) {
 529                 cv_wait(&vd->vdev_initialize_io_cv,
 530                     &vd->vdev_initialize_io_lock);
 531         }
 532         mutex_exit(&vd->vdev_initialize_io_lock);
 533
 534         range_tree_destroy(vd->vdev_initialize_tree);
 535         vdev_initialize_block_free(deadbeef);
 536         vd->vdev_initialize_tree = NULL;
 537
 538         mutex_enter(&vd->vdev_initialize_lock);
 539         if (!vd->vdev_initialize_exit_wanted && vdev_writeable(vd)) {
 540                 vdev_initialize_change_state(vd, VDEV_INITIALIZE_COMPLETE);
 541         }
 542         ASSERT(vd->vdev_initialize_thread != NULL ||
 543             vd->vdev_initialize_inflight == 0);
 544
 545         /*
 546          * Drop the vdev_initialize_lock while we sync out the
 547          * txg since it's possible that a device might be trying to
 548          * come online and must check to see if it needs to restart an
 549          * initialization. That thread will be holding the spa_config_lock
 550          * which would prevent the txg_wait_synced from completing.
 551          */
 552         mutex_exit(&vd->vdev_initialize_lock);
 553         txg_wait_synced(spa_get_dsl(spa), 0);
 554         mutex_enter(&vd->vdev_initialize_lock);
 555
 556         vd->vdev_initialize_thread = NULL;
 557         cv_broadcast(&vd->vdev_initialize_cv);
 558         mutex_exit(&vd->vdev_initialize_lock);
 559
 560         thread_exit();
 561 }
 562
 563 /*
 564  * Initiates a device. Caller must hold vdev_initialize_lock.
 565  * Device must be a leaf and not already be initializing.
 566  */
 567 void
 568 vdev_initialize(vdev_t *vd)
 569 {
 570         ASSERT(MUTEX_HELD(&vd->vdev_initialize_lock));
 571         ASSERT(vd->vdev_ops->vdev_op_leaf);
 572         ASSERT(vdev_is_concrete(vd));
 573         ASSERT3P(vd->vdev_initialize_thread, ==, NULL);
 574         ASSERT(!vd->vdev_detached);
 575         ASSERT(!vd->vdev_initialize_exit_wanted);
 576         ASSERT(!vd->vdev_top->vdev_removing);
 577
 578         vdev_initialize_change_state(vd, VDEV_INITIALIZE_ACTIVE);
 579         vd->vdev_initialize_thread = thread_create(NULL, 0,
 580             vdev_initialize_thread, vd, 0, &p0, TS_RUN, maxclsyspri);
 581 }
 582
 583 /*
 584  * Wait for the initialize thread to be terminated (cancelled or stopped).
 585  */
 586 static void
 587 vdev_initialize_stop_wait_impl(vdev_t *vd)
 588 {
 589         ASSERT(MUTEX_HELD(&vd->vdev_initialize_lock));
 590
 591         while (vd->vdev_initialize_thread != NULL)
 592                 cv_wait(&vd->vdev_initialize_cv, &vd->vdev_initialize_lock);
 593
 594         ASSERT3P(vd->vdev_initialize_thread, ==, NULL);
 595         vd->vdev_initialize_exit_wanted = B_FALSE;
 596 }
 597
 598 /*
 599  * Wait for vdev initialize threads which were either to cleanly exit.
 600  */
 601 void
 602 vdev_initialize_stop_wait(spa_t *spa, list_t *vd_list)
 603 {
 604         vdev_t *vd;
 605
 606         ASSERT(MUTEX_HELD(&spa_namespace_lock));
 607
 608         while ((vd = list_remove_head(vd_list)) != NULL) {
 609                 mutex_enter(&vd->vdev_initialize_lock);
 610                 vdev_initialize_stop_wait_impl(vd);
 611                 mutex_exit(&vd->vdev_initialize_lock);
 612         }
 613 }
 614
 615 /*
 616  * Stop initializing a device, with the resultant initializing state being
 617  * tgt_state.  For blocking behavior pass NULL for vd_list.  Otherwise, when
 618  * a list_t is provided the stopping vdev is inserted in to the list.  Callers
 619  * are then required to call vdev_initialize_stop_wait() to block for all the
 620  * initialization threads to exit.  The caller must hold vdev_initialize_lock
 621  * and must not be writing to the spa config, as the initializing thread may
 622  * try to enter the config as a reader before exiting.
 623  */
 624 void
 625 vdev_initialize_stop(vdev_t *vd, vdev_initializing_state_t tgt_state,
 626     list_t *vd_list)
 627 {
 628         ASSERT(!spa_config_held(vd->vdev_spa, SCL_CONFIG|SCL_STATE, RW_WRITER));
 629         ASSERT(MUTEX_HELD(&vd->vdev_initialize_lock));
 630         ASSERT(vd->vdev_ops->vdev_op_leaf);
 631         ASSERT(vdev_is_concrete(vd));
 632
 633         /*
 634          * Allow cancel requests to proceed even if the initialize thread
 635          * has stopped.
 636          */
 637         if (vd->vdev_initialize_thread == NULL &&
 638             tgt_state != VDEV_INITIALIZE_CANCELED) {
 639                 return;
 640         }
 641
 642         vdev_initialize_change_state(vd, tgt_state);
 643         vd->vdev_initialize_exit_wanted = B_TRUE;
 644
 645         if (vd_list == NULL) {
 646                 vdev_initialize_stop_wait_impl(vd);
 647         } else {
 648                 ASSERT(MUTEX_HELD(&spa_namespace_lock));
 649                 list_insert_tail(vd_list, vd);
 650         }
 651 }
 652
 653 static void
 654 vdev_initialize_stop_all_impl(vdev_t *vd, vdev_initializing_state_t tgt_state,
 655     list_t *vd_list)
 656 {
 657         if (vd->vdev_ops->vdev_op_leaf && vdev_is_concrete(vd)) {
 658                 mutex_enter(&vd->vdev_initialize_lock);
 659                 vdev_initialize_stop(vd, tgt_state, vd_list);
 660                 mutex_exit(&vd->vdev_initialize_lock);
 661                 return;
 662         }
 663
 664         for (uint64_t i = 0; i < vd->vdev_children; i++) {
 665                 vdev_initialize_stop_all_impl(vd->vdev_child[i], tgt_state,
 666                     vd_list);
 667         }
 668 }
 669
 670 /*
 671  * Convenience function to stop initializing of a vdev tree and set all
 672  * initialize thread pointers to NULL.
 673  */
 674 void
 675 vdev_initialize_stop_all(vdev_t *vd, vdev_initializing_state_t tgt_state)
 676 {
 677         spa_t *spa = vd->vdev_spa;
 678         list_t vd_list;
 679
 680         ASSERT(MUTEX_HELD(&spa_namespace_lock));
 681
 682         list_create(&vd_list, sizeof (vdev_t),
 683             offsetof(vdev_t, vdev_initialize_node));
 684
 685         vdev_initialize_stop_all_impl(vd, tgt_state, &vd_list);
 686         vdev_initialize_stop_wait(spa, &vd_list);
 687
 688         if (vd->vdev_spa->spa_sync_on) {
 689                 /* Make sure that our state has been synced to disk */
 690                 txg_wait_synced(spa_get_dsl(vd->vdev_spa), 0);
 691         }
 692
 693         list_destroy(&vd_list);
 694 }
 695
 696 void
 697 vdev_initialize_restart(vdev_t *vd)
 698 {
 699         ASSERT(MUTEX_HELD(&spa_namespace_lock));
 700         ASSERT(!spa_config_held(vd->vdev_spa, SCL_ALL, RW_WRITER));
 701
 702         if (vd->vdev_leaf_zap != 0) {
 703                 mutex_enter(&vd->vdev_initialize_lock);
 704                 uint64_t initialize_state = VDEV_INITIALIZE_NONE;
 705                 int err = zap_lookup(vd->vdev_spa->spa_meta_objset,
 706                     vd->vdev_leaf_zap, VDEV_LEAF_ZAP_INITIALIZE_STATE,
 707                     sizeof (initialize_state), 1, &initialize_state);
 708                 ASSERT(err == 0 || err == ENOENT);
 709                 vd->vdev_initialize_state = initialize_state;
 710
 711                 uint64_t timestamp = 0;
 712                 err = zap_lookup(vd->vdev_spa->spa_meta_objset,
 713                     vd->vdev_leaf_zap, VDEV_LEAF_ZAP_INITIALIZE_ACTION_TIME,
 714                     sizeof (timestamp), 1, &timestamp);
 715                 ASSERT(err == 0 || err == ENOENT);
 716                 vd->vdev_initialize_action_time = timestamp;
 717
 718                 if (vd->vdev_initialize_state == VDEV_INITIALIZE_SUSPENDED ||
 719                     vd->vdev_offline) {
 720                         /* load progress for reporting, but don't resume */
 721                         VERIFY0(vdev_initialize_load(vd));
 722                 } else if (vd->vdev_initialize_state ==
 723                     VDEV_INITIALIZE_ACTIVE && vdev_writeable(vd) &&
 724                     !vd->vdev_top->vdev_removing &&
 725                     vd->vdev_initialize_thread == NULL) {
 726                         vdev_initialize(vd);
 727                 }
 728
 729                 mutex_exit(&vd->vdev_initialize_lock);
 730         }
 731
 732         for (uint64_t i = 0; i < vd->vdev_children; i++) {
 733                 vdev_initialize_restart(vd->vdev_child[i]);
 734         }
 735 }
 736
 737 EXPORT_SYMBOL(vdev_initialize);
 738 EXPORT_SYMBOL(vdev_initialize_stop);
 739 EXPORT_SYMBOL(vdev_initialize_stop_all);
 740 EXPORT_SYMBOL(vdev_initialize_stop_wait);
 741 EXPORT_SYMBOL(vdev_initialize_restart);
 742
 743 /* BEGIN CSTYLED */
 744 ZFS_MODULE_PARAM(zfs, zfs_, initialize_value, ULONG, ZMOD_RW,
 745         "Value written during zpool initialize");
 746
 747 ZFS_MODULE_PARAM(zfs, zfs_, initialize_chunk_size, ULONG, ZMOD_RW,
 748         "Size in bytes of writes by zpool initialize");
 749 /* END CSTYLED */